Re: [PATCH v3 5/7] tracing: Add 'hist' event trigger command

[Tom Zanussi <tom.zanussi@linux.intel.com>]

On Sat, 2015-04-04 at 08:36 -0700, Alexei Starovoitov wrote:
> On Fri, Apr 3, 2015 at 8:51 AM, Tom Zanussi <tom.zanussi@linux.intel.com> wrote:
> > +static struct hist_trigger_entry *
> > +tracing_map_insert(struct tracing_map *map, void *key)
> > +{
> > +       u32 idx, key_hash, test_key;
> > +
> > +       key_hash = jhash(key, map->key_size, 0);
> > +       idx = key_hash >> (32 - (map->map_bits + 1));
> > +
> > +       while (1) {
> > +               idx &= (map->map_size - 1);
> > +               test_key = map->map[idx].key;
> > +
> > +               if (test_key && test_key == key_hash &&
> > +                   keys_match(key, map->map[idx].val->key, map->key_size))
> > +                       return map->map[idx].val;
> > +
> > +               if (!test_key && !cmpxchg(&map->map[idx].key, 0, key_hash)) {
> > +                       struct hist_trigger_entry *entry;
> > +
> > +                       entry = hist_trigger_entry_create(map);
> > +                       if (!entry)
> > +                               break;
> > +                       memcpy(entry->key, key, map->key_size);
> > +                       map->map[idx].val = entry;
> > +
> > +                       return map->map[idx].val;
> > +               }
> 
> There is obvious race here, since algorithm is not implemented correctly.
> Partially inserted key/value is not handled.

You're right, thanks for pointing it out.

I guess adding '&& map->map[idx].val && keys_match(...' would fix the
problem, but would result in duplicate entries for the same key, which I
wanted to avoid.

My original code didn't even have that keys_match() check, which is only
there for the rare possibility of a key collision - I originally just
set a flag in that case, which was printed out with the run results.  In
all my testing I only ever saw it once, so I know it happens very
rarely, but it does happen, and the flag lets the user know and to maybe
try again (my second run of the same test didn't show a collision).  I
thought of actually tracking the collisions as well, since they're so
rare, but didn't see that it was worth the bother.  Using a 64-bit hash
would make it even more remote, but again, is it necessary in practice?

So I think there are several solutions - I'll probably just keep things
simple and go back to a collision flag..

Tom